Prerequisites

Before you begin, confirm that:

• You must have the ROLE_AGENT_AUTHOR on the target tenant. For more information about how to get this role, see Assign Agent Builder roles to users.
• At least one user on the tenant must have the ROLE_AGENT_AUTHOR role (PUBLISH permission). Publishing requires an approver with PUBLISH permission.
• The tools you plan to use must be in the tenant-approved tool catalog.

Step 1: Enter agent and prompt information

In this step, define the agent and create its system prompt. You can either enter the information manually or use Agent Blueprint to generate it from your requirements.

Enter agent and prompt information manually

Use this path when you are building a new agent manually from scratch.

1. Open AgentFlow and go to Agent Builder.
2. Select + Build new agent.

   
   
   

3. Under AGENT DETAILS, enter a name, description, and any relevant tags. These values appear in Discover Agents after the agent is published.

   
   
   

   Use the following rules when filling in the agent details fields:

   • Agent name: Letters, numbers, spaces, and _ - ( ) . & ` are allowed.
   • Describe your agent: Maximum 100 characters. The symbols < and > are not allowed.
   • Tags: Letters, numbers, and hyphens only. Separate multiple tags with commas.

   If a field contains an invalid value, an inline error appears below it. Correct the value before continuing.
4. Write the System prompt. The system prompt defines the agent's:
   1. Identity
   2. Objectives
   3. Tool usage rules
   4. Workflow
   5. Guardrails
   6. Output format
   7. Error handling behavior.

   
   
   

   For more information about structuring an effective system prompt, see AgentFlow system prompt guidelines.

5. Select Save & continue.

   After the agent is created, the Agent ID appears in the AGENT DETAILS section. Select the copy icon next to the Agent ID to copy it to your clipboard.

Generate a system prompt using Agent Blueprint

Follow these steps when you want AI to generate a first draft of the system prompt and select the required tools based on your requirements:

1. Open Agentflow and select Agent Builder.
2. Select + Build new agent and select Build new agent from scratch.
3. Select Generate Agent Blueprint.
4. Enter the required details in the following fields:
   • Agent Name: Enter a clear, specific name that describes what the agent does.
   • Short Description: Enter a brief summary of the main purpose of the agent and the benefit it provides to users.
   • Describe your agent: Describe the role of the agents, the tasks the agent must perform, and the type of responses you need.
   • Intended users: Specify the users or roles the agent is intended for
   • Describe use cases: List the main scenarios or business tasks the agent should support.
   • Constraints: Enter any limits or rules the agent must follow, such as read-only access, restricted tools, or response boundaries.
   • Notes: Enter any additional instructions, preferences, or context that helps improve the generated prompt.
5. Select Generate.

   Agent Builder generates a system prompt draft based on the information you entered. On completion, the Generate agent blueprint pop-up appears.

6. Review the generated system prompt and verify the details.
7. Select See assumptions made to review the inputs used to generate the system prompt.
8. Select Next.
9. Review the list of MCP tools selected for the agent.

   Agent Builder automatically selects tools based on your requirements. You can add or remove tools as required.

   Each tool includes a confidence score of High, Medium, or Low.

10. On reviewing all the details of the agent, select Done.

    The Build an agent page appears.

11. Select Save & continue.

    After the agent is created, the Agent ID appears in the AGENT DETAILS section. Select the copy icon next to the Agent ID to copy it to your clipboard.

To continue, see Step 3: Test & publish the agent

Step 2: Tools & capabilities

1. [Optional] Enable Web search.
   Note: Enabling Web search during agent creation is not sufficient. End users must also enable Web search in their own settings before they can use this capability at runtime.
2. Enable server and select each of the tools you want to add. The selected tools form the agent's allowlist — the agent cannot call tools outside this selection.
   Note: If you proceed without selecting any tools, Agent Builder displays a warning before you can submit for review. You can still proceed, but the agent will have no approved actions available at runtime.

   
   
   

3. Select Save & continue.

Step 3: Test & publish the agent

1. Enter a prompt and review the streaming response.
   Note: If a violation is detected in the system prompt, the test run is blocked. Revise the prompt before testing again.
2. Review the run summary after each test. It shows latency, tokens used, and tool calls for that run.
3. Iterate on the prompt and tool selection as needed.

   
   
   

4. Select Submit for review.
5. Confirm the submission.

   A Confirm submit for review dialog appears. Select Submit for review to proceed, or Back to return to the test panel.

   The system automatically scans the system prompt. Two outcomes are possible:

   • If no violations are detected, the agent transitions to Pending review state.
   • If a violation is detected, the request is auto-blocked. You receive a notification with the scan result. Revise the system prompt and submit again when ready.

   If you need to make changes after submitting, you can withdraw the pending request at any time. Withdrawing returns the agent to Draft state and unlocks the editor.

6. If this agent already has a published version, that version remains live in Discover Agents while the new version is under review.

Update a published agent

Use this path when you want to revise an agent that is already published. An Agent Author (for their own agents) or an Agent Admin can initiate this. The published version remains live and available to users throughout the editing and review process.

Version history is maintained across all draft and published versions of the agent.

Result

The publish request is now in the reviewer queue. For more information about what each state means, see Agent states in Agent Builder.

For more information about what happens next, see Review and act on a publish request.

Validation steps

• Confirm the agent shows Pending review status in Agent Builder.